QCSubmit discussion meeting 2024-02-13

XFF dataset submission

Lily Wang

• Already has two errors somehow for Torsion Drive, only provides traceback for Opt--look into that?

How to get workers running

Jeffrey Wagner

• Recording here, requires OpenFF login

• Jeff is in charge of getting workers up, he will teach us how to do it

• Getting QCFractal workers on a Kubernetes cluster

• Kubernetes is like SLURM but makes Docker images instead of running on a machine somewhere, kind of like a clean new machine so you can install new things and can’t mess anything up

• Need deployment.yaml and manager.yaml, and authentication file for Nautilus

  • Nautilus is preemptible time on supercomputers all over, but very strict about banning people

  • These files contain an access token to modify QCArchive--keep it safe and don’t accidentally modify QCArchive

• Readme file has lots of commands to submit, install software, etc

  • Have a “secret” in manager.yaml which allows us to store QCArchive/authentication stuff

  • “Pod” is an actively running image--one image can make many pods

  • QCFractal worker will get started--knows # processors/RAM, knows programs installed, URL to talk to QCArchive

    • manager.yaml has QCA credentials, needs those to be able to do jobs

    • executors controls number of processors, etc

  • deployment.yaml is like batch script for talking to cluster

    • name for internal use

    • replicas is how many jobs--start small, but once you get the hang of it can increase, change for the right job size

    • selector, template, spec are all computer specific, obtained by talking to sys admin; prioritize getting access to more computers vs not killing jobs since QCA is fault-tolerant

    • container is our docker hub

      • resources--must match between this and manager.yaml. may need to increase CPU or memory at times, make sure to change both

      • volume mounts has QCA authentication secret from earlier

      • env:

        • only allow it to use half the allocated memory because it tends to go over

• LW: what is the easiest way to get into trouble?

  • JW: underutilizing (e.g requesting more than you’re actually using)--they’ll yell at you and at jeff

  • LW: Is there a particular utilization threshhold that we have to stick to?

    • JW: doesn’t remember a specific threshold, seems like other people have wide variety of utilization, will need to keep an eye on it

• LW: with CPU/memory--can we ask for more?

  • JW: usually reach out to admin if we go over 500 CPUs, usually they just say to do whatever you want, you can ask for whatever and they put you in a prioritization queue

  • JW: make sure to periodically check in--when things get resubmitted they (e.g.) can do conda updates, could cause an issue and cause it to crash over and over, make sure that doesn’t happen

• Once you see that the job is working normally, you can increase the number of replicas

  • LW: Do workers kill themselves once they’re out of jobs, or do you have to manually delete them?

    • JW: No, you have to manually delete them, so once many jobs are done you need to keep an eye on it to make sure the utilization is high

    • JW: Can scale it down as well as up, but doesn’t always get ones that aren’t running

• Commands not in README:

  • kubectl get deployment openff-qca-qm

  • kubectl logs openff-qca-qm-xxxxxx….

  • kubectl scale --replicas=y -f deployment.yaml

• To do: obtain Nautilus credentials? https://docs.nationalresearchplatform.org/userdocs/start/quickstart/

  • Next meeting

• https://grafana.nrp-nautilus.io/dashboards Dashboard to monitor CPU usage

  • https://grafana.nrp-nautilus.io/d/85a562078cdf77779eaa1add43ccec1e/kubernetes-compute-resources-namespace-pods?orgId=1&refresh=10s&var-datasource=default&var-cluster=&var-namespace=openforcefield Per-pod dashboard

  • Need to monitor this every few hours while running, if you don’t and the utilization gets too low they might shut down all your jobs

• LW: is there a resubmission mechanism?

  • JW: repo error cycling resubmits jobs that errored out and changes them to waiting

• Downside of PRP and upside of Lilac is that we don’t get many large memory workers, problematic for large molecules

  • Admins don’t like requesting large amount for some molecules but not using it for the occasional normal sized molecule

  • Can use high-memory tag on github if we need to